An mt requires reading only 400ish numbers to work out its internal state and exactly predict its output thereafter. its also biased, which puts bias into any simulation using it.

one of the best performance hwrngs atm comes in the raspberry pi.

one of the worse is that used by (old versions of)ms excel. which im still not sure if they fixed the high probability of creating extreme values.

if you need performance one of the best algorithmic pseudo rngs is in the rabbit stream cipher.

____ quantum random number generator would scare me. whats to stop the nsa from entangling it with their own hardware and controling its output (only half joking)